Claude Opus 5.5 brings lower API prices and strong coding results. Anthropic claims 40% lower costs on typical workloads, but savings depend on token use and caching. We examine the benchmarks, safety measures and what developers should check before switching.
Anthropic released Claude Opus 5.5 on September 22, 2026. It’s the first model in the new Claude 5.5 family, and the pitch is straightforward: performance around the level of its higher-end Fable 5.1 on most work, at a lower running cost than the previous Opus 5.
That is more interesting than another claim that a model topped a leaderboard. Coding agents and research workflows can burn through a surprising number of tokens. If a capable model can finish the same job for less money, people can use it for work they might otherwise consider too expensive.
There is a catch, though. Anthropic’s headline figure of 40% lower cost comes from its tests of typical workloads at default settings. It is not a promise that every user’s bill will fall by 40%.
Claude Opus 5.5 pricing: What actually changed
The listed API prices are $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cache reads fell more sharply, from $0.50 to $0.20 per million tokens. That matters for agents that repeatedly reuse a long conversation, instructions, or project context.
Anthropic says the saving also comes from completing many tasks in fewer steps. But token use depends on the task and the effort setting. In Artificial Analysis’s independent evaluation, Opus 5.5 at maximum effort actually produced more output tokens than Opus 5 and cost about the same per task. Both findings can be true: Anthropic’s typical default-setting workloads and Artificial Analysis’s maximum-effort test measure different things. The number to watch is what your completed task costs.
The basic specifications are familiar. Opus 5.5 has a one-million-token context window, a standard maximum output of 128,000 tokens, and a June 2026 knowledge cutoff. It takes text and images as input and produces text. The Claude API model ID is claude-opus-5-5.
Adaptive thinking is always on. You can’t switch it off, but you can choose an effort level from low through max to balance cost, speed, and quality. The default is medium. Anthropic also offers a faster mode at twice the standard input and output token prices, currently described in its API documentation as a research preview.
Claude Opus 5.5 benchmarks: What the tests show
The numbers are strong, especially for coding and knowledge work. In Anthropic’s published comparison, Opus 5.5 scored 66.4% on Terminal-Bench 4.0, against 55.8% for Fable 5.1, 52.3% for Opus 5, and 57.9% for GPT-6 Astra. It scored 54.4% on the main FrontierCode v1.1 split and 57.8% on CursorBench 4.0.
On GDPval-AA v2.1, which evaluates work across 44 occupations, Anthropic reported an Elo score of 1,846. Fable 5.1 scored 1,735 and Opus 5 scored 1,708 in that comparison.
| Model | API input / 1M tokens | API output / 1M tokens | Terminal-Bench 4.0 | GDPval-AA v2.1 |
|---|---|---|---|---|
| Opus 5.5 | $4 | $20 | 66.4% | 1,846 Elo |
| Opus 5 | $5 | $25 | 52.3% | 1,708 Elo |
| Fable 5.1 | $10 | $50 | 55.8% | 1,735 Elo |
| Sonnet 5 | $2 | $10 | — | — |
Artificial Analysis independently placed Opus 5.5 first on its Intelligence Index at launch, with a score of 58 at maximum effort. That lends weight to the claim that this is a leading model. It also shows why benchmark numbers need context: Artificial Analysis measured 59.6% on Terminal-Bench 4.0 in its own setup, roughly level with its leading GPT-6 Astra result. That is different from Anthropic’s 66.4%, so I would not treat either figure as the model’s universal success rate.
Opus 5.5 doesn’t win everything, either. In Anthropic’s table, GPT-6 Astra was slightly ahead on AutomationBench for business workflows and further ahead on a science-agent benchmark. Different tests reward different strengths; a few points on a leaderboard should not make a purchasing decision for you.
What early users are seeing
The real-work examples are promising, though they are still launch-stage evidence. Anthropic says an early tester completed a 680,000-line code migration in less than a day. In Anthropic’s web-app load-time tests, Opus 5.5 succeeded in 39 of 40 attempts, while Opus 5 made smaller improvements that sometimes changed the app’s behaviour.
Those examples are worth paying attention to. They are also accounts published by Anthropic, not proof that every team can hand a large migration to an agent and come back to finished, reviewed code.
Early user testimonials describe agents that stay on task longer, take fewer steps on some jobs, and produce clearer writing and code comments. That sounds useful. I would still want to see it on my own repository or workflow before calling it a dependable upgrade. Ambiguous requirements, long refactors, and changes with real business consequences still need a person to check the result.
Safety and the bigger picture
This release follows Anthropic CEO Dario Amodei’s call to pace frontier AI development so that safety work can keep up with new capabilities. Anthropic says external groups including METR and Frontier Design tested Opus 5.5 before launch. It also says the model performed better than its previous models on its internal automated behavioural audit.
That is useful information, but it doesn’t settle how the model will behave in every real deployment. Anthropic acknowledges limits to pre-release evaluations and applies safeguards that can route some sensitive requests to other models. For teams building agents, those operating rules matter alongside raw model capability.
Who should care
If you already pay for Opus-class models to run coding agents or long research workflows, Opus 5.5 deserves a proper test. Its listed prices are lower than Opus 5’s, and both Anthropic’s results and independent testing point to strong capability.
If you use Sonnet for everyday work, the decision is less obvious. Claude Sonnet 5’s listed input and output prices are half those of Opus 5.5. A more capable model is only cheaper overall if it completes enough work, saves enough review time, or avoids enough retries to justify the difference.
Developers should also check the migration guide before swapping the model string. Applications that disable thinking, force a particular tool call, or use certain older computer-use integrations may need code changes. A simple integration could be quick to update; an established agent system deserves testing.
Opus 5.5 is available through Claude’s API and on AWS, Google Cloud, and Microsoft’s platform. Anthropic also increased five-hour usage limits for several Claude subscription plans at launch.
The honest take
Opus 5.5 looks less like a dramatic new category of AI and more like a practical improvement to a model people already use for serious work. It has strong early results, lower listed prices than Opus 5, and a credible independent top ranking at launch.
The part I would watch most closely is cost per successful outcome. The 40% saving may hold for some agent workloads and miss for others, especially when effort settings and output length change. The same goes for reliability: a benchmark lead is encouraging, but a clean result on your own messy task is worth more.
For now, Opus 5.5 is worth a serious look if you run sustained coding or knowledge-work agents. Test it on a handful of real jobs, count the review and retries, and then look at the bill. That is where this release will prove its value.